Small SBC Computers

I wrote about an Orange PI ZeroW 4G in my last post. I made the argument that computing is basically free at this point. This is a technical truth, not a practical one. The problem is that they are all bare bones little SBC boards. They need productization. Let me, however, make my argument:

With a few exceptions, computers are already too big. A person sitting in front of a computer can do what ever they do, easily without taxing the computer at all. That’s the same for things like web servers and business logic servers. Here’s the interesting part and where VMware was brilliant: virtual machines!!

With these computers being so big, so overly powerful, we have to sub-divide them into littler machines to make use of our bigger machines. The whole of “cloud” is based on virtual machines being carved out of larger machines, but does that make economic sense in the long run? There are architectural limitations of these big machines. The CPU chips have only so many cores. They can address only so much memory and have to resort to NUMA. The cost of these systems increases at a non-linear curve as the core and memory density increases.

Lets look at some numbers, an 32 core AMD CPU at time of writing is about $600. A 4 core ARM system is about $34 add a USB power supply for $6 and call the price $40/each and 8 of those would cost about $320, but they have the whole “computer” around them and would have 32G of RAM, while the AMD would need a motherboard and RAM. 32 Gig of RAM is about $80. A motherboard costs about $150. A power supply is another $150. That’s a total of $980 vs $320. I’m ignoring the cost of physical packaging right now for the sake of the discussion.

At the moment, the Orange PI is good for low level operations, you’d still want something heftier for day to day work, but as a web server, storage server, computational server or any other thing that you’d spin up a VM for, they would be fine. That’s kind of the problem….

Your big expensive computer is actually too limited to be a practical VM server. With only 32 cores how many VMs can you actually run? After some point, you will find you need a bigger computer or yet another computer. If, however, you create a “farm” of these little computers, and treat them like they were “VMs” now you have a cost effective scalable compute farm. This is nothing more than a physical packaging and software management problem.

Going back to our calculations, assuming the packaging issues are handled and the software management system is usable. You can double your compute farm power for about $320. You can’t do that with your big computer.

As we know, the growth of bigger computers is slowing down, while the density and power of the tiny SBCs is still growing strong. You can get them with more RAM and better I/O. VMs are nice and useful, but SBC compute nodes will be the future. You can even use containers on these SBCs to add to their flexibility.

It may take a while, but we are already seeing raspberry PI compute boards and off-the-shelf clusters getting built. Its still not “enterprise ready,” but it will be. They use less power, are less of a capital investment, and, let’s be honest, almost disposable. Just by stacking more and more of them, your server farm can grow at a near linear cost per compute node.